The Happiness Score
You might know what makes you happy, but what makes humans happy in general? Is it good food? Spending time with a loved one? Asai et al asked workers on Mturk to answer the question ‘what made you happy in the past 24 hours (or alternatively, the past 3 months)’ and what resulted is a juicy dataset of 100,000 happy sentences.
While this dataset was created keeping NLP in mind, I wanted to know if we could understand something more about this unique emotion called ‘happiness’ through written text. The authors of the happydb paper also shared scores, the Valence, Arousal and Dominance scores (VAD) for each sentence. The VAD model was sourced from the paper by Bradley and Lang, 1994 and will be used for all further analysis here.
What does Valence, Arousal and Dominance have to do with my happiness?
Valence is positive or negative affectivity, whereas arousal measures how calming or exciting the information is. The Semantic Differential Scale devised by Mehrabian and Russell (1974) is a widely used instrument for assessing the 3-dimensional structure of 18 bipolar adjective pairs that are each rated along a 9-point scale.
They conceived pleasure as a continuum ranging from extreme pain or unhappiness to extreme happiness and used adjectives such as happy-unhappy, pleased-annoyed, and satisfied unsatisfied to define a person’s level of pleasure. Arousal was conceived as a mental activity describing the state of feeling along a single dimension ranging from sleep to frantic excitement and linked to adjectives such as stimulated-relaxed, excited-calm and wide awake-sleepy to define arousal. Dominance was related to feelings of control and the extent to which an individual feels restricted in his behavior. To define the degree of dominance Mehrabian and Russell used a continuum ranging from dominance to submissiveness with adjectives such as controlling, influential and autonomous.
Thus, a certain emotional state can be quantified and levels of emotions can be measured.
1. Giving induces more happiness than taking
If your parents ever taught you (which I hope they did) to share and help others, there is a proven reason for it. Analysis of people’s experiences show that activities like giving, sharing and helping induces greater levels of pleasure than taking or receiving.
2. You remember events that have a greater emotional impact
A paper by Cahill and McGaugh refer Godard’s research that electrical stimulation of Amygdaloid complex stimulates arousal and activates the EEG which heightens memory. Our data set proves this research for the case of positive stimuli. A statistically significant difference is found in arousal levels for the 24 hour and 3 month reflection period. The 3 month reflection period has a higher arousal because of which they recollect it to the date of when the question was asked.
3. Expression of happiness is not just a positive statement
I wondered how expressing happiness is different from a positive tweet on Twitter. As can be seen from the plots below, recollection of happiness is a far more emotional statement, as can be seen from high VAD scores, than a positive self-declared statement on twitter.
If you look more closely, tweets have higher arousal, almost as much as happiness statements, because users wouldn’t have tweeted something if it did not ctach their attention in the first place. There is an interesting divide between the two datasets for dominance, showing that the respondents to the happydb dataset seem to experience more control and confidence than the twitter users. Valence doesn’t show as much a gap because both statements are positive/happy ones.
What can I do with these findings?
Sensing the emotional intent behind a sentence can have various business use cases. It can be used to identify fake reviews on sites like Yelp and TripAdviser, because the score will give the authenticity of a positive review/statement. It can be used to enhance conversion of texts to emojis. VAD scores play a huge role in marketing - successful advertisements induce high valence, low arousal. This dataset can be used to train models and choose the right adjectives and verbs to suit the product segment.